Generative artificial intelligence offers new ways to support personalized learning, yet many existing applications emphasize content generation without fully connecting learning evidence, instructional intervention, and subsequent evaluation. Adopting a design and development research approach, this study designs a multi-agent personalized learning support system with mastery learning theory as its instructional framework. Within the bounded context of a learning space, specialized agents share data and coordinate tasks to diagnose mastery from formative assessment evidence, generate corrective or enrichment support, and prepare a second parallel formative assessment after teacher review. This process creates an interpretable and traceable closed loop of personalized learning support. The study further derives five design principles and develops a learning-evidence model, a multi-agent collaboration architecture, personalized learning packages, and a staged evaluation framework. Its principal contribution is the translation of mastery learning into an executable and auditable multi-agent workflow, offering design guidance for personalized learning support systems. The proposed design and prototype establish feasibility claims only; their educational effectiveness remains to be examined empirically.